One of the most basic goals in vision research is to understand how visual information is represented at the level of the retinal output cells, the ganglion cells, as these cells provide all the information about the visual world the brain receives. These cells are made up of many different classes, each with its own sensitivities to visual stimuli, and each producing its own set of signals. How these cells work together to collectively form visual representations has been a long-standing critical question - one whose answer is needed both for basic science (for understanding fundamentals of visual processing) and for applied science (for developing algorithms to drive visual prosthetics). We recently developed a tool for addressing this and use it for both these purposes. Briefly, the tool is a retinal input/output model. It differs from other models in that it's effective on a broad range of image statistics, including those of white noise, gratings, natural scenes (landscapes, faces, etc.) With the model we can make rapid advances on both goals. Our Specific Aims are the following: Aim 1 is to test hypotheses about the roles of the different ganglion cell classes in representing visual images. We use the model to build the hypotheses, and then electrophysiology (multi-electrode recording) experiments to test them. Aim 2 focuses on the development of a new retinal prosthetic strategy. We used the model combined with optogenetics to develop a system that can produce normal retinal output, that is, it can make blind, degenerated retinas produce normal firing patterns to a broad range of stimuli, including spatiotemporally varying natural scenes. Here we will develop and expand the method, specifically, so that it is effective not just for ganglion cells but also for bipolar cells, as thse are the two major stimulation targets for retinal prosthetics, and each has its own strengths. This approach produces substantially better (near-normal) representation of visual images than existing methods.